A Robust Deep Retinex Decomposition Network Leveraging a Novel Synthetic Dataset for Low-Light Image Enhancement

Retinex-based low-light image enhancement methods decompose an image into illumination and reflectance components, improving the quality of images captured under poor lighting conditions by adjusting the illumination component. However, the ill-posed nature of the Retinex decomposition and the lack of ground truth for decomposition results have led existing methods to rely on various hand-crafted priors to impose additional constraints on the decomposition process, which may compromise the quality and authenticity of the decomposition results. To address this issue, we construct a novel synthetic PRI dataset, consisting of 2,700 rendered and illumination images under different lighting conditions. Leveraging this dataset, we design a robust deep Retinex decomposition method. Comparative experiments with four classical Retinex-based methods on the LoLv1 and PRI datasets demonstrate the superiority of our decomposition method. Additionally, cross-network comparisons and low-light pedestrian detection experiments further validate that our method effectively enhances the low-light image enhancement performance of baseline methods. Our codes and dataset will be publicly available at https://github.com/Kylesesa.